mamba create -n bench -c anaconda -c conda-forge ipython numpy "libblas=*=*mkl" "liblapack=*=*mkl" intel-openmp

• =*=*openblas: uses OpenBLAS; default
• =*=*mkl: uses Intel’s Math Kernel Library (MKL)
• =*=*accelerate: uses Apple’s Accelerate framework with vecLib
• =*=*blis: uses AMD' AOCL-BLIS

TODO:

• how about MKL/Accelerate builds for scipy and sklearn?
• does intel have optimizations or hardware acceleration for Pandas

import numpy as np
def test(size):
		left = np.random.random((size, size))
		right = np.random.random((size, size))
		return np.sum(left @ right)
		
%timeit test(2048)

# AMD 5850U Note: single channel
# BLIS (prioritizes one core at max frequency)
## 365 ms ± 2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
# MKL (prioritizes multi-core but physical only)
## 202 ms ± 6.42 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
# MKL but dual channel
## 190 ms ± 5.66 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

# AMD 2950X quad channel
# BLIS
## 649 ms ± 2.45 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
# MKL (also multi-core physical)
## 151 ms ± 2.14 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

# Mac Intel i7-8750H
# MKL
## 132 ms ± 2.57 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

mamba create -n py "python>=3.13" ipython numpy numba powerlaw pyspark matplotlib seaborn pandas networkx python-build "libblas=*=*mkl" "liblapack=*=*mkl" intel-openmp ruff pydantic jupyterlab scikit-image scikit-learn simdjson cython

Matrix Multiplication (2048x2048) on Apple M1 (4+4)

• OpenBLAS: 94.7 ms ± 2.34 ms per loop
• Accelerate: 68.6 ms ± 321 µs per loop
• Notes: OpenBLAS uses all available cores, while Accelerate only dedicates one performance core. Accelerate still ends up better with less CPU usage.

Matrix Multiplication (2048x2048) on Mac Intel i7 (6)

• OpenBLAS: 96.5 ms ± 6.72 ms per loop
• MKL: 74.6 ms ± 3.35 ms per loop
• Accelerate: 99 ms ± 4.18 ms per loop